The Effectiveness of Image Augmentation in Deep Learning Networks for Detecting COVID-19: A Geometric Transformation...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Science

Background:

• Chest X-rays offer an accessible and affordable method for early COVID-19 pneumonia detection.
• Deep learning models show promise for COVID-19 detection on X-rays but require large datasets to prevent overfitting.
• Image augmentation is commonly used to enhance deep learning model performance by increasing training data.

Purpose of the Study:

• To investigate the impact of geometric augmentations on deep learning algorithms for COVID-19 detection using chest X-rays.
• To compare the performance of 17 deep learning models with and without various geometric augmentation techniques.

Main Methods:

• Evaluated 17 deep learning algorithms on chest X-ray datasets for COVID-19 detection.
• Compared model performance with and without four different geometric augmentation strategies.
• Analyzed the influence of augmentation on detection accuracy, dataset diversity, methodology, and network size.

Main Results:

• Contrary to expectations, removing geometric augmentations improved the Matthews correlation coefficient (MCC) for all 17 models.
• The MCC without augmentation (0.51) surpassed results from four tested geometric augmentation methods (ranging from 0.44 to 0.49).
• Retraining a recent deep learning model without augmentation significantly increased detection accuracy (p-value = 2.23 × 10⁻³⁷).

Conclusions:

• Geometric augmentation may not be beneficial and can potentially hinder deep learning model performance for COVID-19 detection on X-rays.
• The findings suggest a need to re-evaluate and potentially remove geometric augmentation steps in current deep learning algorithms for COVID-19 screening.
• Provides clinical perspectives for developing more robust COVID-19 X-ray-based detection systems.